Search papers, labs, and topics across Lattice.
B and
1
0
3
LLMs can be sped up by over 2x without sacrificing accuracy, by compressing the input and predicting multiple output tokens at once using a unified framework.
